Somewhere in the fine print of a Microsoft partner-center announcement, a number died. Twenty-five thousand messages for two hundred dollars — a tidy, comprehensible price — quietly expired. What replaced it was not a new price. It was a unit of measurement: four thousand credits per user per month, default-on, opt-out.
I did the arithmetic the way I always do, the way I did it in 2017 when I audited forty-plus whitepapers and watched three tokenomics models collapse under a Python simulation. Microsoft's own worked example: one hundred users, four hundred thousand credits, thirty-two hundred dollars a month. That is thirty-two dollars per user per month of potential overage, stacked on top of a seat fee that has itself become almost symbolic — five dollars at Salesforce, twenty at Anthropic. The seat is dead. Long live the meter. And the meter, it turns out, is a currency nobody agreed on, issued by eight vendors who each define it differently, redeemable nowhere, and inflatable at will. That is the real story here, and it is a story crypto has been telling itself for a decade without realizing the enterprise world was about to relive it.
Here is the landscape, stripped of marketing. Eight enterprise AI agent platforms now publish — or refuse to publish — pricing in five incompatible units: seats, credits, actions, tokens, and conversations. Salesforce Agentforce sells a five-dollar user license, then charges five hundred dollars per hundred thousand Flex Credits, and defines a single action as twenty credits. Do the math and an action is ten cents. Microsoft Copilot Studio sells two hundred dollars per twenty-five thousand credits — eight-tenths of a cent each. Anthropic's enterprise tier anchors at twenty dollars a seat per month with usage billed separately at API rates. Meta's entry point is free, monetized at two dollars per million tokens on WhatsApp replies. xAI publishes its API rate card — two dollars and six dollars per million tokens — then goes silent on enterprise deals. Google and ServiceNow publish nothing at all; third-party estimates float between twenty-one and sixty dollars and between seventy and two hundred-plus dollars respectively, never confirmed on any official page.
Four of eight hide enterprise pricing entirely. The other four expose units that cannot be compared to each other. This is not a market. It is eight parallel markets wearing the same word. I have seen this movie before. In 2020, at a Berlin hackathon, I helped build a crude bot to track liquidity-mining rewards across Uniswap and Aave, and the first thing we discovered was that APY meant something different on every protocol. The number was real. The unit was fiction. We spent more time normalizing definitions than writing code — and that was the moment I stopped trusting any single protocol's dashboard as a source of truth.
Let me start with the reverse-engineering, because the numbers are the only honest thing in this entire mess. Meta charges two dollars per million tokens. The source material pegs a WhatsApp reply at four to five cents. Divide. That is twenty thousand to twenty-five thousand tokens per single message. For context, an ordinary chat turn runs five hundred to two thousand. An agent burns an order of magnitude more because the entire apparatus rides inside the context window — system prompt, tool definitions, multi-turn history, RAG retrieval, and the tool's return payload, all re-sent, all metered, every single turn. This is the technical fact underneath the pricing chaos: a single agent interaction is a high-variance, high-magnitude event, and no pre-AI billing unit survives contact with it. Microsoft said so out loud when it retired the message-based SKU, conceding that consumption depends on how the agent is designed and used, and that a message no longer corresponds to a fixed price. That is a vendor admitting the same word can mean ten times the cost depending on architecture.
Now the currencies, because this is where the crypto reader should sit up straight. Microsoft's credit is eight-tenths of a cent. Salesforce's is half a cent. That is a sixty percent spread — an arbitrageur's dream, except you cannot trade them. Credits are non-convertible, non-transferable, and adjustable at the issuer's sole discretion. They are not a unit of account. They are vendor-issued private currencies with an undisclosed inflation schedule — structurally identical to airline miles, which is precisely the mechanism crypto spent a decade trying to escape. A private currency with a unilateral inflation lever and no disclosure obligation is not a pricing innovation. It is a quiet transfer of value from buyer to issuer, executed one consumption-rate change at a time.
And here is the sharpest edge of the whole thing. Salesforce defines an action as twenty credits. But what is an action? Check order status is an action. Issue a refund and notify the customer is also an action. The vendor controls the boundary. Whoever defines the unit controls the price without ever touching the sticker. That is the metering layer, and it is the actual battlefield — not the rate card, which is theater, but the definition of the thing being metered. Compare the price archaeology: Salesforce still displays a two-dollar-per-conversation card, marked in small type as valid only for existing Agentforce Conversations customers. A buyer who searches, who reads a third-party review, who screenshots the page, will find a number that no longer applies. If a conversation averages twenty actions, the old and new prices converge. If it averages five, the action model costs four times more. That conversion ratio has never been published. It is the single number that determines whether an existing customer's bill quietly quadruples, and it is invisible — a broken oracle in a market that desperately needs one.
I built a tokenomics calculator once because I was tired of watching people argue about numbers nobody could decompose. This is the same disease. The vendor publishes a rate. The vendor hides the conversion. The buyer cannot reconstruct the bill. Now scale that into an enterprise procurement process and watch it break. A traditional SaaS budget is seats times price times discount, with error controllable to within five percent. An agent budget is a function of usage intensity, task complexity, context length, and human-review ratio — and at least two of those variables are completely unknown at the moment of purchase. This is a paradigm fracture in financial planning, not a problem that more phone calls to the sales rep can fix. When four of eight platforms hide enterprise pricing and the other four speak incompatible units, a request for proposal cannot even be designed. There is no common scoring dimension. Procurement teams fall back on the three subjective levers they always fall back on: relationship, brand, and proof-of-concept feel. The math leaves the room, and vibes walk in.
Then there is the default. Microsoft's CSP channel is moving new Copilot Business subscriptions onto usage-based billing by default — opt-out, not opt-in — with a four-thousand-credit-per-user monthly cap, and the United States is not on the exemption list. Behavioral economics has a name for this: the default effect. Most customers never change a default. Which means a large population of enterprises will enter metered billing without ever making a conscious decision about it. The most consequential pricing change in enterprise AI this year is not a number. It is a checkbox that ships pre-checked. And the channel mechanics compound it: Microsoft is simultaneously unbundling the base subscription, swapping SKU identifiers, and flipping metering on by default, with the date slipping from early November to December. That slip is itself a signal — end of sale is not end of life, and the postponement tells you the channel side was not ready. Every small managed service provider in that ecosystem now faces billing-system rebuilds and awkward customer conversations, and those costs flow downhill to the end client.
And what is the metric of success? The source buries the most damning line in its sixth paragraph: a thirty-thousand-seat reference deployment, cost undisclosed, measured by adoption, not productivity. Read that twice. Enterprises are buying AI agents and grading them on how much they are used, not how much they save. Metered billing running underneath an adoption KPI is a machine engineered to reward overconsumption. It is a slot machine that pays out in usage graphs, and the house always keeps the margin.
Now the tiering, because the tiering is the tell. Tier one is Microsoft and Salesforce — ecosystems and channels — who publish rate cards and hide the variables. Tier one, second flavor, is the model labs: OpenAI's enterprise tier is fully closed and contact-sales only, while Anthropic anchors at twenty dollars a seat and bills usage on top. Tier two is Google and ServiceNow, fully quote-gated, leaning on existing enterprise relationships. Tier three is the challengers, xAI and Meta, who publish prices precisely because they must. The correlation is exact and it runs backward from intuition: the stronger the vendor, the less transparent the price. Opacity is not a market failure. It is a moat. It blocks price wars, protects margin, and shifts competition onto the axes the leader already owns — ecosystem, channel, integration depth. xAI runs the opposite playbook, transparent at the top of the funnel and closed at the bottom, exactly as AWS did in its early years. Meta's free entry with a two-dollar token price tied to its own channel is price-as-lock-in, and the price is the lock. Google's position is the most awkward of all: no Microsoft channel, no lab-grade model reputation, yet fully closed on pricing — which suggests the opacity there is not a strategy but the symptom of a decision not yet made.
And the psychological anchor: Anthropic's twenty-dollar seat matches ChatGPT Plus almost to the dollar, whispering that enterprise is not expensive. But usage bills separately, and the real cost lands at five to twenty times the seat. It is the cleanest anchoring misdirection in the set, and it works because the anchor is the only number the buyer remembers. Meanwhile the questions that actually matter — is the credit consumption rate locked in the contract? What happens at overage, does the service stop, degrade, or climb a pricing ladder? Do unused credits roll over or evaporate at month end? What is the minimum commit, the real reason quote-gated often means below-some-threshold-and-we-will-not-answer? What discount band applies, given that enterprise list prices and actual deals routinely diverge by thirty to seventy percent? — none of these appear on any page. The contract terms are ten times more important than the list price, and they are the one thing the entire transparency debate refuses to discuss. And can the customer even audit the meter? Can you prove this action truly consumed twenty credits, or that an agent caught in a runaway loop did not burn a month of budget in an afternoon? Without auditability, metering is a black box with a bill attached.
The hidden labor is the part that should terrify anyone who has ever run an audit. Every automated action — the refund, the rebooking, the customer email — needs a human to check it. That review is the load-bearing wall. It is priced at zero: not on the vendor invoice, not in the AI project's ROI, not in employment statistics, where it is filed under existing headcount rather than new AI cost. The total cost of ownership of an AI agent is systematically underestimated because the most expensive component — human verification — is invisible to the accounting system that is supposed to measure it. This is the same blind spot I found covering NFT ownership in 2021: everyone priced the JPEG and nobody priced the labor of believing in it. And the metering layer is becoming the control point the way CloudWatch and the billing system — not EC2 itself — became AWS's moat. Salesforce pushes credits. Microsoft pushes credits. Meta and xAI push tokens. This is a standards war dressed as a pricing complaint, and almost no one is naming it as such. Whoever defines the meter defines the industry.
Here is where I part company with the source's own headline. The title says nobody will show you. But the article's own body concedes the opposite: the problem is not that anyone is hiding the price. So which is it? The honest answer is that the transparency framing is the wrong frame, and it flatters the buyer into thinking the fix is more disclosure. It is not. The vendors are not hiding prices. They are hiding conversion rates, default settings, and unit definitions — and disclosure of a rate card does nothing about any of the three. You can publish every price and still leave the buyer unable to budget, because the variable that determines the bill is not the price. It is the definition of an action, the default state of a checkbox, and the ratio between old and new units. There is a second blind spot the source misses entirely: the beneficiaries of opacity. When pricing is unmeasurable and metering is baroque, enterprises do not walk away. They hire Accenture. They hire Deloitte. They hire the local systems integrator. Every layer of pricing confusion is a consulting engagement waiting to be scoped — and that interest chain stays invisible in a piece that reads like consumer advocacy. And the antitrust question goes unasked: if Microsoft bundles Copilot with Azure consumption and flips metering on by default, that is a tying and self-preferencing problem with a European regulator's name written on it. The opacity story is the symptom. Bundling is the structure.
The seat is gone. The meter rules. And the meter is a private currency issued by eight banks that answer to no one, denominated in units they define, convertible at rates they refuse to publish. Crypto spent a decade learning the hard way that a unit of account only works when its issuer cannot silently inflate it. The enterprise AI market is walking into the same trap at a thousand times the scale, one pre-checked checkbox at a time. So here is the question I will be asking on every earnings call for the next two years: when the metering layer becomes the trust layer, will the standard be written by the vendors who profit from opacity — or by the auditors who can prove what an action actually cost? The meter is a ledger too, and someone has to read it. Where the code meets the chaotic human heart, the ledger is still blank. Rewriting the ledger, one story at a time.